Final The Growing Longevity Gap between Rich and Poor and Its Impact on Redistribution through Social Security Barry Bosworth and Gary Burtless THE BROOKINGS INSTITUTION and Kan Zhang Gianattasio * GEORGE WASHINGTON UNIVERSITY October 18, 2016 * The authors are, respectively, senior fellows in the Economic Studies program at the Brookings Institution, Washington, D.C., and a graduate student in health policy at the George Washington University. We are indebted to Mattan Alalouf and Eric Koepcke for excellent research assistance. The authors gratefully acknowledge funding for this research from the Alfred P. Sloan Foundation through its Working Longer program and from the Social Security Administration as part of the Retirement Research Consortium. The views are solely our own and do not represent those of Brookings, the Sloan Foundation, or the Social Security Administration.
The Growing Longevity Gap between Rich and Poor and Its Impact on Redistribution through Social Security by Barry P. Bosworth, Gary Burtless, and Kan Zhang Gianattasio 1 THE BROOKINGS INSTITUTION Washington, DC October 18, 2016 Abstract This paper uses interview data from the Survey of Income and Program Participation (SIPP) combined with Social Security Administration records on worker earnings, pension benefits, and mortality to determine the extent of widening differences in U.S. life expectancy by socioeconomic status (SES). We construct alternative measures of SES using respondents educational attainment and average Social-Security-covered earnings in their prime working ages, 41-50. We find large mortality rate differences among Americans past age 50 when they are ranked by their SES. More importantly, the differences have grown significantly in recent years, a finding confirmed using both measures of SES and alternative procedures to sort respondents into high and low SES groups. Comparing men born in 1920 and 1940, for example, the estimated gain in life expectancy at age 50 for men in the bottom one-tenth of the mid-career income distribution was only 1.7 years versus a gain of 8.7 years among men in the top tenth of the income distribution. Among women in the same two birth cohorts, the expected longevity gains in bottom and top income deciles were 0.0 years and 6.4 years, respectively. It matters little whether we use education or mid-career earnings to measure SES. The conclusions are also robust across samples that include or exclude workers who claim disability benefits. We find secular changes in differential mortality to be very large, but the influence of widening mortality differences on the length of time Americans collect Social Security benefits is damped by the fact that low SES workers and spouses tend to claim pensions at younger ages while high SES workers are more likely to postpone retirement and benefit claiming. The changes in relative mortality rates across the income distribution has offset a growing percentage of the progressivity built into the Social Security benefit formula. We find, however, that the resulting pattern of lifetime benefits nonetheless remains progressive, even taking account of the widening mortality differentials in recent years. 1 Bosworth and Burtless are Senior Fellows in the Economic Studies program at the Brookings Institution in Washington, DC. Gianattasio is a graduate student in health policy at George Washington University, Washington, DC. -1-
The Growing Longevity Gap between Rich and Poor and Its Impact on Redistribution through Social Security by Barry P. Bosworth, Gary Burtless, and Kan Zhang Gianattasio 1 October 18, 2016 U.S. LIFE SPANS are unequal, and part of the inequality is linked to Americans social and economic position. A long-standing literature shows that there are significant differences in life expectancy between people with high and low socioeconomic status as measured by income and educational attainment. New studies show that the longevity differential has widened in the United States over the past three decades, reversing an earlier trend toward narrower differentials (Waldron 2007, 2013; NAS 2015; Bosworth, Burtless and Zhang 2016; Chetty et al. 2016). A few studies suggest that much of the recent increase in expected life spans is concentrated among those with above-average incomes, and that life expectancy may be roughly constant or even declining for Americans with lower status. Recent trends in differential mortality raise profound questions about the equity of oldage pension formulas. The Social Security retirement-worker pension provides a basic benefit at the normal retirement age, known as the Primary Insurance Amount or PIA. The formula for this pension is based on a worker s average lifetime earnings and is highly redistributive. It provides a more generous replacement rate for pensioners with low earnings compared with workers who have high career earnings. This kind of redistribution helps compensate low-wage workers for their shorter expected life spans. Retired workers actual monthly benefits are determined by their PIA and the actuarial factors used to adjust the monthly pension to reflect early or late benefit claiming. Workers who claim benefits at the earliest entitlement age, 62, receive reduced benefits; workers who delay claiming benefits until the latest claiming age, 70, receive a monthly payment that is about three-quarters greater than a benefit claimed at age 62. Since the early 1990s the average U.S. retirement age has trended upward, and there has also been a trend toward later claiming of Social Security benefits (Bosworth and Burtless 2010; Bosworth, 1 We gratefully acknowledge the vital research contributions of Mattan Alalouf and Eric Koepcke. This paper summarizes some of the research findings reported in Bosworth, Burtless and Zhang, Later Retirement, Inequality in Old Age, and the Growing Gap in Longevity between Rich and Poor (Washington: Brookings, 2016). We also acknowledge the generous research support provided by the Alfred P. Sloan Foundation and the Social Security Administration. -2-
Burtless and Zhang 2016). Workers who delay benefit claiming receive bigger monthly pensions as a result. If these workers earned better-than-average wages during their careers, the delay in benefit claiming increases the gap between their monthly Social Security benefits and the monthly benefits received by lower wage workers, who tend to retire at younger ages and claim benefits sooner after age 62. Differences in mortality mean that high-wage workers can expect to collect benefits longer than low-wage workers who claim benefits at the same age. Because gains in expected life spans are increasingly concentrated among high-wage workers, the proportional gap in lifetime benefits between high- and low-wage pensioners has grown in recent years. A common suggestion to deal with funding shortfalls in Social Security and Medicare is to lift the age of eligibility for benefits. This policy makes sense if gains in expected life spans are enjoyed equally by rich and poor. It seems less equitable to ask low-wage workers to wait longer for retirement benefits when a disproportionate share of gains in life expectancy is enjoyed by the affluent. In view of the changing relationship between workers average lifetime earnings and their chances of surviving into late old age, how can we recalibrate the PIA formula and the actuarial adjustment for delayed benefit claiming to protect the interests of low-wage workers? This paper reports on research that expands the base of knowledge with which to answer this question. The remainder of the paper is divided into two parts. The first section presents our results on the growing mortality differential between Americans based on their average Social-Securitycovered earnings. We organized a large data file containing information from the Census Bureau s Survey of Income and Program Participation (SIPP) matched to Social Security lifetime earnings records and Social Security mortality records. These data permit us to analyze determinants of mortality within a large sample of SIPP respondents born between 1910 and 1950 over the period from 1984 through 2012. Besides obtaining access to these confidential data, the most challenging part of the research is devising measures of socioeconomic status that permit us to make evenhanded comparisons between generations born over a 40-year time span. This is a challenge because the measures of annual earnings contained in the Social Security Administration (SSA) files were subject to different reporting limits during the ages we use to estimate average earnings for successive birth cohorts. In our analyses of these data, we used two methods for dealing with the limitations of the earnings data. -3-
The second section of the paper uses our estimates of the widening mortality rate differential between higher and lower income workers to estimate the impact of growing lifespan inequality on the lifetime redistribution that takes place through Social Security. The rising mortality gap between the affluent and poor Researchers have found evidence of widening in mortality differentials by social and economic status in a number of recent studies. Singh and Siahpush (2006) offer evidence on changes in differential mortality using county-level information from the decennial censuses. They construct county-level indexes of SES linked to death records by location. Meara, Richards and Cutler (2008) and Olshansky and others (2012) analyze death certificate data, using educational attainment as a measure of status, and find a sharp rise in inequality. Waldron (2007) uses administrative records containing information on career earnings and age at death to establish a similar pattern for men covered by Social Security. Bosworth, Burtless, and Zhang (2015) and NAS (2015) use longitudinal data from the Health and Retirement Study (HRS) combined with Social Security earnings and death records to estimate the relationship between mortality and lifetime earnings and other indicators of social and economic status. Both studies find significantly faster gains in life expectancy among top Social Security earners compared with low earners in recent decades. Chetty et al. (2016) combine information from income tax records and Social Security death records covering the period from 1999 to 2014 to determine trends in life expectancy by percentile of the national income distribution and, within geographical regions, by quartile of the income distribution. In the time period they examine, life expectancy increased 2.3 years among men and 2.9 years among women in the top 5 percent of the U.S. income distribution. In contrast, among those in the bottom 5 percent of the distribution life expectancy increased just 0.3 years among men and 0.04 years among women. The SIPP sample. Our recent research extends our earlier analysis of the matched HRS files by estimating trends in mortality using SIPP longitudinal survey files that were matched to earnings, benefits, and mortality records drawn from the Social Security administrative files. Our dataset contains records for SIPP respondents born after 1910 who were interviewed in the 1984, 1993, 1996, 2001, and 2004 panels. 2 Our principal analysis focuses on men and women 2 For a description of the SIPP samples, sampling methodology, and interview methods see URL = http://www.census.gov/programs-surveys/sipp/methodology/organizing-principles.html. -4-
born before 1951. We were able to successfully match about 80 percent of the SIPP respondents to their corresponding Social Security earnings and death records, yielding a total sample of 41,000 men and 45,000 women (Table 1). We could also match about 95 percent of respondents who were married, with spouse present at the time of the SIPP interview to their spouse s Social Security record. Note, however, that the SIPP interviews covered only about two and a half years of each respondent s career. We do not have post-interview information about respondents later marriage partners. The mortality rate in our analysis sample was 37 percent in the case of men and 30 percent for women. To perform our detailed analyses of mortality rates by age we created a person-year dataset. Each respondent enters the sample in the year of attaining age 50 or the year corresponding to his or her initial SIPP interview, whichever occurs later. Respondents remain in the sample until the year they die or until 2012, the last year for which we have reliable information about mortality. Our final dataset contains 487,000 person-year observations for men and 573,000 person-year observations for women. Indicators of socio-economic status. Analysts have used four main markers of socioeconomic status to indicate individuals position within the social hierarchy: educational attainment; income or earnings; occupation; and wealth. In our recent analysis, we focused on two of these indicators, educational attainment and mid-career Social-Security-covered earnings. This summary focuses mainly on our findings based on mid-career earnings. One reason is that findings based on these two indicators produced broadly similar results. This is reassuring. Educational attainment and Social-Security-covered earnings each have some shortcomings as indicators of socio-economic status for respondents born over a 40-year time span, as we argue below. The fact that our analyses using the two indicators produce the same finding with respect to widening mortality differentials tends to confirm the inference that the differential is in fact growing. Many studies of the link between socio-economic status and mortality have used education as the principal indicator of status. Its measurement is straightforward and reasonably accurate in most household surveys. Education is ordinarily determined by early adulthood, well before we begin measuring mortality rates in middle age and late adulthood. Educational attainment does have some limitations, however. Many studies of the effect of SES on mortality use an absolute rather than a relative measure of attainment. This can be problematic when -5-
schooling attainment has risen strongly across successive generations. Such is the case for Americans born during the four decades after 1910. In the 1962 Current Population Survey, 58 percent of the men who were between 48 and 52 years old (and born between 1910 and 1914) reported they had not completed high school; just 9 percent reported they had completed college. In the 1998 Current Population Survey, only 14 percent of 48-52 year-old men (born between 1946 and 1950) reported they had failed to complete high school; 33 percent reported they had obtained a college degree. Clearly, lack of a high school diploma was an indicator of much deeper social disadvantage for 48-52 year-old men in 1998 than it was in 1962. Completion of college was a more marked indicator of social and economic advantage in 1962 than it was in 1998. If we find that failure to complete high school is associated with a much bigger increase in mortality among men born in 1946-1950 compared with those born in 1910-1914, we can hardly be surprised. Men who failed to complete high school represented a much smaller and more disadvantaged population in 1998 compared with 1962. In our analysis we deal with this problem by converting SIPP respondents educational attainment reports into number of years of schooling and then normalizing each person s years of schooling relative to the average of the educational attainment of their immediately surrounding birth cohorts (people born within two years before or after the person s birth year). In essence we are measuring education as the deviation of the person s own attainment compared with that of the average attainment of people born within two years of the person s birth. The calculations were done separately for men and women, effectively eliminating any trend in our measure of relative education across successive birth cohorts. 3 Mid-career earnings as an indicator of socio-economic status. Several early studies of the effect of SES on mortality used current income as an indicator of status because it was the only available measure of income in the household survey used by the analyst. Current income has some problems as an indicator of SES because of its sensitivity to adverse health shocks or other transitory, income-reducing events. The availability of Social Security earnings records makes it possible to construct an average of workers past earnings, a measure we shall refer to as mid-career earnings. This measure of SES avoids many of the problems caused by using a 3 Under this adjustment procedure, the normalized level of schooling has the same mean number of years for each 5-year birth cohort. -6-
single year s income. A 10-year average of mid-career earnings dilutes the role of transitory influences and may come close to the concept of permanent income. Our use of average earnings in mid-career also reduces, though it does not eliminate, the potential for reverse causation flowing from health to income. The quality and limitations of the Social Security earnings data have varied over the years. Until 1978, the Social Security Administration maintained its own earnings records based on quarterly reports of employers. In 1978 SSA switched to reliance on annual earnings information collected by the IRS. Between 1951 and 1977 the earnings data were limited to covered earnings up to the annual taxable wage ceiling. Unfortunately, the ceiling wage was not regularly adjusted to reflect changes in the distribution of earnings. The ceiling wage was only 3 percent above the economy-wide average earnings level in 1965 but 69 percent above average earnings in 1977. There are two broad approaches to dealing with the limitations of the Social Security earnings records before 1978. The first is to use information in the Social Security earnings files to predict annual earnings for workers whose annual earnings are above the taxable earnings ceiling. The second approach is to use only information on workers reported annual earnings below a maximum percentile level. The maximum percentile is selected to correspond with the capped earnings amount in the calendar year with the lowest taxable earnings ceiling relative to the earnings distribution. In our longer research paper, we estimated the relation between mortality rates and workers mid-career earnings using both of these procedures. Both yielded broadly similar results, described below. This summary of our research findings focuses on results obtained when we rely on actual reported earnings in those years when a worker s earnings are below the taxable wage ceiling and on imputed earnings in those calendar years when the worker s reported Social-Security-covered earnings are equal to the maximum taxable amount. To make our predictions we impute workers earnings above the taxable wage ceiling using information on the quarter in which the worker s earnings reached the maximum taxed amount. For workers who reached the ceiling with 4 quarters of reported earnings, the imputed annual total wage was set to 1.14 times the taxable maximum. For those with 3 quarters, we assigned an imputed amount equal to 1.53 times the taxable maximum. For those with two quarters, the imputed ratio was 2.36. For those who reached the ceiling in the 1 st quarter, the -7-
imputed ratio was set at 5 times the taxable maximum. 4 The annual earnings data available since the early 1980s has the major advantage of providing measures of earnings in excess of the taxable wage ceiling. In addition, it includes earnings from both Social Security covered and uncovered jobs. We cap the annual earnings distribution at the 98 th percentile to reduce the impact on our results of a few very large values in the post-1977 data. 5 We created a measure of mid-career average earnings by first deflating each worker s nominal annual earnings using the SSA average wage index with a base year of 2005. This procedure largely eliminates the influence of secular economy-wide wage growth on our measure of workers annual earnings. For each individual worker, we calculated mid-career average earnings as the mean real nonzero earnings amount when the worker was between ages 41 and 50. 6 The resulting mean values of workers mid-career average earnings are shown separately for men and women by birth year in Figure 1. These earnings estimates raise some of the same issues already mentioned in our discussion of educational indicators of SES. Because women have been increasingly likely to be employed and to earn higher relative wages in recent birth cohorts, their career earnings have increased compared with those of men. Meanwhile, the average (indexed) wage of men has declined for the youngest birth cohorts. Note that the economy-wide earnings index includes the annual wages of all workers in a given calendar year, rather than only those of workers between 41 and 50. It follows that our indexed estimates of 4 The adjustment ratios were originally derived for a report to SSA (Toder et al. 1999). Class intervals were set under an assumption of steady earnings throughout the year, and the class means were derived from the distribution of wages in various reports of the Current Population Survey. Less than one percent of the workers in the sample reached the taxable maximum in the first quarter. A similar methodology was also used more recently in Cristia (2009) and Kopczuk, Saez, and Song (2010). Additional problems with the changeover to W-2 records in 1978-80 led us to use an interpolation of individuals earnings above the taxable ceiling between 1977 and 1981. No adjustment could be made for the self-employed who were above the taxable wage ceiling as they file on an annual basis. 5 Even after our adjustments, the pre-1977 data are not fully compatible with the later years because of bunching of imputed earnings after adjusting for the quarter in which individuals reach the taxable wage ceiling. 6 In other words, calendar years in which a worker had no reported earnings were excluded in calculating the worker s average earnings. The computation of mid-career average earnings is adapted from Waldron (2007). As she noted, the reliance solely on years with of nonzero earnings excludes some low-wage workers who have very poor health in middle age. However, by excluding zero earnings years, our measure probably gives a more reliable indicator of workers potential earnings in years not affected by unemployment or severe health problems. -8-
mid-career earnings will be affected by changes in the age distribution of the overall work force as well as the average wage of 41-50 year-olds relative to other earners. To eliminate any secular drift in our estimates of average earnings across birth cohorts we employed an adjustment similar to the one we used to convert individuals educational attainment into a normalized measure of education. In particular, we calculated the mean indexed earnings of each overlapping 5-year birth cohort, and we then normalized individual earnings levels within a 5-year birth cohort so that the mean normalized value was equal to that of the 1938-1942 birth cohort. Thus, an individual s mid-career earnings is measured relative to the average mid-career earnings of people the same sex born within two years before or two years after the worker s own birth year. 7 We calculated these estimates of workers normalized earnings separately for men and women. Majorities of men and women in our sample were married at the time of the SIPP interview. For married women, the woman s normalized mid-career earnings are often a poor indicator of socioeconomic status. Married women in older cohorts frequently did not work outside the home or earned very low annual wages, especially during years they were rearing children. For this reason, it makes sense to use a household-based, rather than an individualbased, measure of earnings to classify the socio-economic status of members of married couples. Bearing this in mind, we combined normalized husband and wife earnings as our main incomebased measure of household-level SES. We defined equivalized household earnings for individuals with a spouse as the sum of the two mid-career normalized earnings amounts divided by the square root of two. 8 For respondents who did not have a spouse, we used the worker s own mid-career earnings. Our measurement procedure requires us to exclude from the analysis all SIPP respondents who were single and did not have positive Social-Security-covered earnings between ages 41 and 50. It also requires exclusion of married couples where neither spouse had positive covered earnings. 7 Under this adjustment procedure, the normalized level of mid-career earnings has the same mean value for each 5-year birth cohort and is equal to the mean for the birth cohorts born from 1938 to 1942. 8 This is a common procedure for converting the total income of a two-person family into the equivalent income of a one-person household. Economists estimate equivalent incomes by determining the change in expenditure that is required to hold living standards constant when a household gets larger or smaller. One popular adjustment, which we use here, assumes that a household s spending requirements increase in proportion to the square root of the number of household members. -9-
As noted, we also used an alternative procedure to minimize the impact reporting changes in the historical records of workers Social-Security-covered earnings. Our alternative procedure uses only information on workers reported SSA earnings up to a maximum percentile level. In the case of men the maximum percentile is less than the median male earnings level in each year. 9 We selected the maximum percentile so that the earnings we counted in each calendar year would be measured in a consistent way, regardless of whether the maximum taxable amount in the year was high or low in relation to the earnings distribution in that year. Obviously, the alternative method does not permit us to distinguish between the earnings of male workers who earned average and well-above-average earnings, but it does give a consistent method for distinguishing low-earnings men from men with average- or above-average incomes. In the remainder of this paper, we will use information about the full range of earnings reported in the Social Security records, including plausible imputations of earnings above the maximum taxable earnings amount. Other variables. In addition to the SES indicators just described, a number of other personal characteristics are known to affect mortality rates. Two of these are race or ethnicity and marital status. The SIPP interview also includes a self-reported measure of respondents health status. This indicator identifies one of the channels through which variations in SES may influence mortality. The SIPP contains a health indicator that can range from a value of 1 for respondents in excellent health to 5 for those in poor health. We also have an indicator showing whether the respondent was ever disabled. For SIPP sample members, this information can be derived from the individual s Social Security benefit record, which shows whether the person ever claimed Disability Insurance. Estimation of mortality risks. Our statistical analysis is based on a logit regression of mortality risk that takes the form: 9 In the mid-1960s, the maximum Social-Security-taxable earnings amount was attained at the 31 st percentile of the earnings distribution of 41-to-50 year-old men. Therefore, this was the maximum level of annual earnings we used in order to classify male workers as low wage earners. Women s reported earnings were much less affected by the maximum taxable earnings amount. In years when the taxable cap was low relative to economy-wide average wages, in 1951 through the early-1970s, working women s earnings were also comparatively low. As a result, annual wages up through the 80 th earnings percentile of 41-to-50 year-old women were observed, even in the calendar year with the lowest earnings cap relative to the female earnings distribution. -10-
(1) ( h it 1 h it ) = exp(β ij X ijt ), where h it = Pr(Y it = 1 / Y it-1 = 0) is the hazard that person i will die in year t; and X ijt = Vector of determinants of mortality risk. The determinants we include are the person s SES, age, birth year, and categorical variables for race/ethnicity, marital status, and disability. 10 Ages range from 50 to 100. Our indicator of birth year is the person s year of birth minus 1900. The birth year is our basic indicator of cohort effects. We employ two alternative indicators of SES (mid-career earnings and educational attainment), social and economic indicators that are potentially linked to differential mortality. For each of these measures we also include the interaction of SES with the birth year in order to estimate the rise or decline in differential mortality across successive birth cohorts. The fact that that the age-specific mortality rate is constrained to increase by a fixed proportional amount at every age and across successive cohorts may impose an excessively severe restriction on the mortality function. We experimented with alternative nonlinear measures of age, but they were never statistically significant. In our longer paper we also show the effect of estimating widening mortality differentials within narrower age groups than the age 50-to-100 age group that is examined here. The logistic regression results for the SIPP sample are displayed in Table 2. Mortality risks are estimated separately by gender. We show three sets of regression results for each sex. The first column reports results from a regression that includes age, birth year, and the two alternative measures of SES (a) our equivalized and normalized indicator of household midcareer earnings; and (b) our normalization of respondents educational attainment. The coefficient on Age indicates a rising probability of death as respondents grow older. The coefficient on Birthyear is negative, indicating that age-specific mortality risks are declining for successive age cohorts. The specification in the first column includes both SES indicators, and their estimated coefficients are negative and highly statistically significant as expected. 11 Thus, 10 In addition, we included a categorical variable for the first calendar year of a respondent s enrollment in the sample, recognizing the fact that respondents were exposed to the risk of dying for less than a full 12 months in that calendar year. 11 We estimated versions of the mortality equations that limited the measure of SES to either career earnings or educational attainment, but the combination of both variables yielded a superior overall statistical relationship without altering the coefficients on the non-ses indicators. -11-
consistent with earlier research we find strong statistical evidence of differential mortality by social and economic status. People with high equivalized earnings or greater educational attainment have lower rates of mortality than those who have less earned income or less schooling. Marital status and past disability are also highly significant predictors of mortality risk. The role of race is more marginal, however, once we include measures of mid-career earnings and educational attainment. Our tests for increasing differential mortality are displayed in columns 2 and 3 (for men) and columns 5 and 6 (for women). The columns show results when we add an interaction between birth year and one of the two measures of SES. While the interaction complicates the interpretation of the other coefficients, a negative coefficient on the interaction term implies that the size of the mortality difference across levels of mid-career earnings or educational attainment is increasing in later birth cohorts. Perhaps surprisingly, there is very little to choose between the two indicators of SES. They both yield very significant negative coefficients on the interaction of the SES indicator with the birth year, implying strongly increasing differential mortality. The measures of overall explanatory power are also virtually identical. The interaction of education with birth year has a larger negative coefficient than the interaction of mid-career earnings with birth year, which may seem to suggest a more pronounced pattern of increasing differential mortality. However, the larger coefficient is due to the much more limited range of variation in the educational attainment variable compared with mid-career earnings. The coefficient estimates for men and women show very similar influences for many of the determinants of mortality. For both sexes there is powerful evidence of increasing differential mortality using either mid-career earnings or educational attainment as an indicator of SES. It is noticeable, however, that marital status has a consistently smaller impact on female mortality. Women s mortality is apparently less adversely affected by living in an unmarried state. The implications of our logit estimates for mortality rates at ages 65 and 75 are displayed in Figures 2 and 3. The results in Figure 2 show predicted mortality rates of men born in successive cohorts between 1920 and 1945, while the results in Figure 3 show comparable predictions for women in the same birth cohorts. The predictions for men are based on the parameter estimates displayed in column 2 of Table 2, and those for women are based on the parameter estimates in column 5 of the same table. In both cases, our estimate of the changing mortality gradient linked to SES is based on the interaction of respondents equivalized mid- -12-
career earnings and their birth year. We derive all the predictions in Figures 2 and 3 for men and women who, with the exception of their mid-career earnings, have average characteristics for respondents in the middle one-fifth of the equivalized mid-career earnings distribution. Each chart contains three lines, and each line displays the trend of age-specific mortality rates for successive birth cohorts at a given income level. The top line in each chart shows the trend in mortality rates for individuals in the bottom one-tenth of the equivalized mid-career earnings distribution. The middle line shows the trend in mortality rates for those in the middle fifth of the mid-career earnings distribution, and the bottom line shows the mortality trend for men or women in the top one-tenth of the mid-career earnings distribution. The charts suggest that, holding other determinants of mortality constant, respondents born in 1920 had similar mortality rates regardless of whether their households mid-career earnings placed them at the top or the bottom of the income distribution. In later cohorts, however, the mortality rate difference that is linked to respondents mid-career earnings widens. For a male 65-year-old born in 1940, for example, a respondent in the top one-tenth of the household mid-career earnings distribution faces a mortality risk that is just 40 percent of the risk faced by a person with the same characteristics who is in the bottom one-tenth of the distribution. In the case of 65-year-old women born in the same year, the mortality risk facing a woman in the top tenth of the income distribution is just 38 percent of that facing a woman in the bottom income decile. The mortality rate declined in later birth cohorts for men at all income levels, but it fell much faster among men with above-average compared with below-average incomes. Among women there is little evidence that women at the bottom of the mid-career earnings distribution saw any mortality improvement at all. Women with average and above-average equivalized household mid-career earnings, however, saw meaningful declines in their mortality rates. Note that our predictions are derived from a sample in which we do not observe all of the age and birth year combinations displayed in the charts. For example, because our mortality data end in 2012, our sample does not contain any 75-year-old men or women who were born in 1940. Consequently, some of our mortality rate calculations are based on out-of-sample predictions. The out-of-sample predications are indicated in the charts by broken rather than solid lines. -13-
Implications for life expectancy. Using the mortality equations in Table 2, we calculated the probability of death at each age between 50 and 100 and cumulated the results for each sample member in order to obtain the probability of survival to each age: (2) S x = S x 1 (1 D x ), where Dx is the expected conditional death rate at age x. To see how life expectancy varies by respondents SES, we ranked all sample members by their normalized mid-career earnings and divided the samples into ten equal-size groups. Within each decile we calculated the mean life expectancy for the men or women in the earnings decile. The distribution was calculated for both individual and equivalized household mid-career earnings, but here we report only the results based on the equivalized household income measure. For women in particular we believe this variable represents the better measure of a person s relative SES. Most of the following analyses focus on simulated expected life spans for the 1920 and 1940 birth cohorts. These two years represent the extremes of the birth years for which we have reasonably complete earnings records and a sizeable number of actual deaths. We performed simulations in which alternative estimates of life expectancy were computed using equations 1 and 2 and the coefficients reported in Table 2 for all people in the SIPP sample described above. To simulate the life expectancy of a population born in 1920, we replaced each sample member s actual birth year with 1920 and calculated their expected remaining life assuming they survived to age 50. To simulate the life expectancy of a population born in 1940 we followed the same procedure, replacing each sample member s actual birth year with 1940. 12 The resulting estimates of life expectancy for the simulated1920 and 1940 birth cohorts are shown in Table 3. 13 The top panel of Table 3 shows simulated life expectancies for men. The results displayed in three columns on the left are based on our specification that uses equivalized midcareer earnings interacted with Birthyear to capture the changing effect of SES on mortality 12 The analysis is based on a simple exercise in which the estimated mortality equation is used to generate a predicted mortality rate for each age from 50 to 100, using the birth years of 1920 and 1940 in turn. 13 Note that the characteristics of the populations represented in Table 3 are those of the entire SIPP sample, not just the members of that sample born in 1920 or 1940. We are attempting to measure the change in life expectancy between 1920 and 1940, and we are estimating that change in a population with the fixed characteristics of our entire estimation sample. -14-
rates. Those in the three columns on the right are based on the specification that uses relative educational attainment interacted with Birthyear to capture the changing effect of SES. The implied differences in life expectancy between the top and bottom SES groups are very large. The simulation results for the 1920 cohort suggest that men in the top decile of mid-career earnings could expect to live 5.0 years longer than men in the bottom decile 79.3 versus 74.3 years. For men born twenty years later in 1940, the simulated improvements in life expectancy added an average of 4.8 years to male life spans. However, the gain in life expectancy was only 1.7 years for men in the lowest earnings decile compared with a gain of 8.7 years for men in the top decile. The gains are thus heavily skewed towards men at the top of the income distribution. When we use education interacted with Birthyear to capture the changing effect of SES (columns 4 6 in Table 3), the increase in average life expectancy is nearly the same, but the differential gains in life expectancy between the top and bottom of the mid-career earnings distribution appear smaller. For men the simulated increase in life expectancy is 3.8 years in the bottom income decile compared with a gain of 5.5 years in the top decile. The gap in longevity gains between the top and bottom deciles in columns 4 6 would appear wider if we ranked men by their educational attainment rather than by their equivalized household mid-career earnings. When we constructed decile rankings based on educational attainment instead of equivalized earnings, the simulated increase in male life expectancy between the 1920 and 1940 birth cohorts was 2.2 years in the bottom one-tenth of men and 6.5 years in the top tenth. We observe a smaller increase in simulated average life expectancy among women between the 1920 and 1940 birth cohorts. The average gain implied by the SIPP results is only 2.7 years. The gains are highly correlated with women s SES, however. When we use equivalized mid-career earnings interacted with Birthyear to capture the changing effect of SES, there is in fact no apparent increase in life expectancy in the lowest income decile. This compares to a gain of 6.4 years in life expectancy for women in the top earnings decile. Because life expectancy gains among women were on average slower than they were among men, there is a noticeable narrowing in the life expectancy gap between women and men. When we use education interacted with Birthyear to capture the changing effect of SES, there is somewhat weaker evidence that the increase in life expectancy is correlated with higher levels of household -15-
earnings. The simulated gain in life expectancy is 2.0 years in the lowest decile and 3.0 years at the top. 14 Figure 4 summarizes the changes in predicted life expectancy for men and women. The top panel of the chart shows changes in male life expectancy by fifths of the equivalized household earnings distribution. The lower panel shows predictions for women. Both charts show a strong upward tilt in favor of higher income people, for men and women born in both 1920 and 1940. As we have already seen, however, the upward tilt is considerably more favorable for those born in 1940 compared with those born in 1920. The column of figures on the right of the chart shows the years of gain in life expectancy between 1920 and 1940 by fifth of the income distribution. The changes in life expectancy between the two birth cohorts within each tenth of the mid-career earnings distribution are graphically displayed in Figure 5. The top line shows longevity improvements among men, while the bottom line shows gains among women. Although life span gains were faster among men than among women, for both sexes the improvements in life expectancy past age 50 were considerably faster at the top of the income distribution than at the bottom. Robustness checks. In our longer paper we present results from an identical specification using data from the HRS interviews matched to Social Security earnings and mortality records. The HRS matched files have some disadvantages compared with the SIPP files. First, the enrolled HRS sample is smaller than the SIPP sample. Second, there was a lower match rate of HRS and SSA earnings records, further reducing the HRS sample size. Finally, the HRS sample was enrolled later than the earliest SIPP samples, giving us evidence on the SES-mortality gradient over a shorter range of years. Although the HRS results were not identical to the ones displayed in Table 2, they strongly confirmed the basic finding reported here: Differential mortality has increased significantly and markedly across successive birth cohorts. The estimated coefficient on the SES interaction with birth year was negative and highly significant. However, in the case of men the size of the coefficient on the interaction between education and birth year was smaller and less significant than it is in the SIPP sample. 14 Again, the gap in life expectancy gains between the top and bottom deciles in columns 4 6 would appear wider if we ranked women by their educational attainment rather than by their equivalized household mid-career earnings. -16-
In another robustness check, we re-estimated the relationship between SIPP respondents mid-career earnings and their age-specific mortality rates using very basic information about mid-career earnings patterns reported in the Social Security Administration files. In particular, we classified men as low earners under a range of definitions that relied solely on reported earnings below the 31 st percentile of the annual male earnings distribution. 15 This information is available (without any imputation) for all men in our SIPP sample, regardless of the calendar years in which the men were between 41 and 50 years old. Similarly, we classified women as low earners based on their mid-career earnings relative to those of other women born in the same year. Further, we classified women as members of low earnings households based on the combination of their own Social-Security-covered earnings and the low earnings status of their spouse. In all cases, the classifications were based solely on observed (rather than imputed) midcareer Social-Security-covered earnings of men and women in the SIPP sample. We then estimated discrete-time logistic models of mortality risk using our binary classification of low earners or low earnings households as indicators of respondents SES. To determine whether the mortality differential between low earners and average or above-average earners increased over time, we included an interaction term between the low earnings indicator and the respondent s year of birth. The results obtained using this alternative methodology strongly confirmed the results displayed in Table 2. For both men and women born between 1910 and 1956 the estimated coefficients showed a statistically significant increase in the mortality rate difference between low earnings workers and workers with average or above-average earnings. Respondents classified as low earners saw noticeably slower reductions in mortality over successive birth cohorts compared with respondents with average or above-average earnings. We found the same pattern among both male and female SIPP respondents, regardless of our classification scheme for identifying low earners or, in the case of women, low earnings households. Furthermore, when we subdivided our overall sample into smaller subsamples restricted to observations in narrower 7- or 9-year age groups, we found statistically significant and meaningfully large increases in the mortality differentials in most of the age groups we analyzed. 15 For mid-career male earners in the mid-1960s, the maximum taxed earnings amount was attained at the 31 st percentile of the earnings distribution. See note 9. -17-
Explanations. The reasons for the increase in differential mortality across SES groups remain uncertain. In particular, it is unclear whether differential access to health care is the main channel through which SES affects mortality, as opposed, for example, to socio-economic differences in behaviors, such as smoking, drinking, and lack of regular exercise, that are linked to early mortality. Using a large sample of adults age 25-64 covering the period of 1970 to 2000, Cutler and others (2010) concluded that behavioral factors, such as smoking and obesity, have strong effects on mortality risk, but they contributed little to explaining the growing disparity in mortality by levels of educational attainment. The HRS includes information on self-reported heath status and some behavioral risks, including alcohol use, smoking, and levels of physical activity. We re-estimated our HRS mortality regressions to include heath status and the three behavioral risk factors. 16 In all cases, health status and the behavioral variables had high statistical significance in predicting mortality, but they had a relatively small effect in reducing the size or statistical significance of the coefficient on the interaction term between the SES indicator and birth year our measure of increasing differential mortality. Including these variables does not serve as a substitute for the SES-birth year interaction. Overall, we interpret these results as showing that a consistent pattern of increasing differential mortality is operating through channels in addition to health status and the behavioral measures. In all of our analyses, disability has a very large and significant positive impact on mortality risk. Yet inclusion or exclusion of disability status has very little impact on the estimated size of the coefficient on the SES-birth year interaction. When we exclude from the estimation sample respondents who received a DI benefit, there is very little effect on the coefficients of the SES indicators or their interaction with respondents year of birth. Thus, our estimates of widening mortality differentials linked to SES are quite robust to alternative methods for dealing with disability. One possibility is that rising income inequality has increased the mortality rate differences between Americans with a high and a low rank in the income distribution. Whatever the longevity advantage conferred by higher income, the fact that the proportional income gap between high and low income Americans has widened over the past three decades may have had 16 Including the number of alcoholic drinks consumed in a typical day averaged across all survey waves, whether individuals ever smoked or were smoking at the time of the last interview, and whether they engaged in vigorous physical activity at least three times per week. -18-